A pixel art converter that lets you choose the colors

Because another entry on your list sits between it and everything in the picture. Nearest-color matching is a competition and only one entry can win a cell, so a color that is close to a color you already have will never be the closest to anything — this is most common when two entries were sampled a few pixels apart from the same gradient. Delete it, or move it far enough away that it owns a region of its own. A zero is a fact about your list, not about the image.

The entries that won the fewest cells in a first full conversion, skipping any you pinned. That ranking matters: dropping by position in the list would delete whichever color you happened to add last, which has nothing to do with whether the picture needs it. After the drop the whole image is converted again against what survived, so the cells that belonged to a removed color are re-matched properly. That second pass is why the counts move when a cap is applied rather than simply redistributing.

It exempts that entry from the cap, so it stays in the match however few cells it wins. It is for the colors that carry meaning rather than area — the single red of a logo, the highlight in an eye, the one warm tone in an otherwise cold picture. Those are exactly the entries a count-based cap removes first, because being small is the whole point of them. Pinning does not force cells onto a color; it only stops the cap taking it away.

Yes, and the list will tell you it was pointless. Two entries a few units apart in sRGB split whatever region they both sit near, so you get two rows in the spend table with roughly half the cells each and a picture that looks the same as it did with one. The matching runs in Oklab, which is close enough to perceptual that two entries you cannot tell apart on screen will also be near-indistinguishable to the search — if you want both to survive, move one until you can see the difference in the swatches.

Between eight and twenty-four for most artwork, and the number matters far less than the spacing. A well-spread twelve beats a badly spread thirty-two, because the thirty-two will have spent half its entries on one crowded part of the range and left the rest of the picture with nothing near it. The spend column is the check: if your top three entries hold more than about 70% of the cells between them, the list is unbalanced rather than too short.

Yes. Cells are filled with the exact value on the list, not with a blend or an approximation of it, so a pixel in the exported PNG samples back to the hex you typed. That is not true of the averaging step before it — the mean color of a cell is computed in linear light and can be any value at all — but that mean is only ever used to decide which of your entries wins. What is painted is always one of yours.

The front page, which does the same conversion against eight fixed palettes and needs no decisions from you. This page exists for the opposite case: you already know what colors the result has to be, either because you are matching an existing piece of art or because you own a specific set of physical materials and cannot buy more of them.